because saving i1 and i2 to their ``designated'' stack slots corrupts unknown
memory in other functions, standard libraries, and worse.
In addition, this has the benefit of improving JIT performance because we
eliminate writing out 4 instructions in CompilationCallback() and 2 loads and 2
stores.
llvm-svn: 7653
2. Handle fp-to-uint conversions directly here instead of relying on
a pre-transformation to replace them with the 2-step conversion.
3. Use size rather than explicitly checking types when deciding what
opcodes to use, wherever possible. This is less error prone (the
bug fix above was not the first time!).
4. Float-to-pointer casts shd now work though this hasn't been tested.
llvm-svn: 7645
* Doxygen-ified comments
* Added capability to make far calls (i.e., beyond 30 bits in CALL instr)
which implies that we need to delete function references that were added by
the call to addFunctionReference() because the actual call instruction is 10
instructions away (thanks to 64-bit address construction)
* Cleaned up code that generates far jumps by using an array+loop
SparcV9CodeEmitter.h:
* Explained more of the side-effects of emitFarCall()
llvm-svn: 7639
Note that some generated operators (like &, | or ^) may
not be supported by the assembler -- but if they've got
this far, it's better to generate them and let the assembler decide.
llvm-svn: 7476
since it is *necessary* for correct code generation. Only optional
transformations belong in the PreOpts pass (which needs to be renamed
from PreSelection to PreOpts).
llvm-svn: 7474
that depends on machine register size.
Moved insertCallerSavingCode() to PhyRegAlloc and
moved isRegVolatile and modifiedByCall to TargetRegInfo: they are all
machine independent. Remove several dead functions.
llvm-svn: 7392
Single and FP double reg types (which share the same reg class).
Now all methods marking/finding unused regs consider the regType
within the reg class, and SparcFloatRegClass specializes this code.
(2) Remove machine-specific regalloc. methods that are no longer needed.
In particular, arguments and return value from a call do not need
machine-specific code for allocation.
(3) Rename TargetRegInfo::getRegType variants to avoid unintentional
overloading when an include file is omitted.
llvm-svn: 7334
now works in instructions which require a 2-bit or 3-bit INTcc code.
Incidentally, that means that the representation of INTcc registers is now the
same in both integer and FP instructions. Thus, code became much simpler and
cleaner.
llvm-svn: 7185
allow, i.e. make a sequence of instructions to enable an indirect call using
jump-and-link and 2 temporary registers (which we save and ultimately restore).
Warning: if the delay slot of a function call is used to do meaningful work and
not just a NOP, this behavior is incorrect. However, the Sparc backend does not
yet utilize the delay slots effectively, so it is not necessary to make an
overly complicated algorithm for something that's not used.
llvm-svn: 7178
* FP double registers are now coded correctly
* Removed function which converted registers based on register types, it was
broken (because regTypes are broken)
llvm-svn: 7175
(1) Cannot use ANDN(ot), ORN, and XORN for boolean ops, only bitwise ops.
(2) Conditional move instructions must distinguish signed and unsigned
condition codes, e.g., MOVLE vs. MOVLEU.
(3) Conditional-move-on-register was using the cond-move-on-cc opcodes,
which produces a valid-looking instruction with bogus registers!
(4) Here's a really cute one: dividing-by-2^k for negative numbers needs to
add 2^k-1 before shifting, not add 1 after shifting. Sadly, these
are the same when k=0 so our poor test case worked fine.
(5) Casting between signed and unsigned values was not correct:
completely reimplemented.
(6) Zero-extension on unsigned values was bogus: I was only doing the
SRL and not the SLLX before it. Don't know WHAT I was thinking!
(7) And the most important class of changes: Sign-extensions on signed values.
Signed values are not sign-extended after ordinary operations,
so they must be sign-extended before the following cases:
-- passing to an external or unknown function
-- returning from a function
-- using as operand 2 of DIV or REM
-- using as either operand of condition-code setting operation
(currently only SUBCC), with smaller than 32-bit operands
Also, a couple of improvements:
(1) Fold cast-to-bool into Not(bool). Need to do this for And, Or, XOR also.
(2) Convert SetCC-Const into a conditional-move-on-register (case 41)
if the constant is 0. This was only being done for branch-on-SetCC-Const
when the branch is folded with the SetCC-Const.
llvm-svn: 7159
(1) An int CC live range must be spilled if there are any interferences,
even if no other "neighbour" in the interf. graph has been allocated
that reg. yet. This is actually true of any class with only one reg!
(2) SparcIntCCRegClass::colorIGNode sets the color even if the LR must
be spilled so that the machine-independent spill code doesn't have to
make the machine-dependent decision of which CC name to use based on
operand type: %xcc or %icc. (These are two halves of the same register.)
(3) LR->isMarkedForSpill() is no longer the same as LR->hasColor().
These should never have been the same, and this is necessary now for #2.
(4) All RDCCR and WRCCR instructions are directly generated with the
phony number for %ccr so that EmitAssembly/EmitBinary doesn't have to
deal with this.
llvm-svn: 7151
correct: empirically, "regType" is wrong for a number of registers. Thus, one
can only rely on the "regClass" to figure out what kind of register one is
dealing with.
This change switches to using only "regClass" and adds a few extra DEBUG() print
statements and a few clean-ups in comments and code, mostly minor.
llvm-svn: 7103
the TableGen descriptions; all unset bits are thus errors.
* As a result, found and fixed instructions where some operands were not
actually assigned into the right portion of the instruction.
llvm-svn: 7074
area to avoid using up precious stack space within the 4095 offset limit
from %fp. Such objects that would themselves live at a large offset
were being put there already so this is a simple change.
llvm-svn: 7066
deletes it, but we were merrily trying to fix the operands of that
instruction anyway! Instead, fix the replacement instruction.
(2) An Improvement: Check for and extract global values in all operands,
not just in known pointer operands. For example, they can occur in
call arguments, and probably other unforeseeable places as well.
This also eliminates the special-case handling of Load and Store.
llvm-svn: 7053
(1) select: Ok to convert a pointer to a float or double.
(2) regalloc: Some MachineInstr* for caller-saving code before a call
were being inserted before and after the call!
(3) Don't insert the caller-saving instructions in the
MachineCodeForInstruction for the Call instruction.
*All* instructions generated by register allocation need to be
recorded in those maps, but it needs to be done uniformly.
llvm-svn: 7051
Move LowerAllocations, PrintFunction, and SymbolStripping passes, and
the corresponding -disable-strip and -d options, over here to the SPARC
target-specific bits of llc. Rename -d to -dump-asm.
tools/llc/Makefile:
Reindent. Add x86 library so that llc compiles again.
tools/llc/llc.cpp:
Remove support for running arbitrary optimization passes. Use opt instead.
Remove LowerAllocations, PrintFunction, and SymbolStripping passes, as noted
above.
Allow user to select a backend (x86 or SPARC); default to guessing from
the endianness/pointer size of the input bytecode file.
Fix typos.
Delete empty .s file and exit with error status if target does not support
static compilation.
llvm-svn: 6776
* BPA and BPN do not take a %cc register as a parameter
* SLL/SRL/SRA{r,i}5 are there for a reason - they are ONLY 32-bit instructions
* Likewise, SLL/SRL/SRAX{r,i}6 are only 64-bit
* Added WRCCR{r,i} opcodes
llvm-svn: 6655
* Removed instruction scheduling as it is too slow to run in a JIT environment
* Removed other passes because they aren't necessary and can slow JIT down
llvm-svn: 6652
Special cases: STFSRx and STXFSRx - they operate on predefined rd=0 or rd=1, and
expect %fsr as the parameter in assembly. They are disabled (since not used)
until an encoding, both for code generation and output, is chosen.
llvm-svn: 6619
Update file comment to contain a bunch of the overview mapping-info
documentation previously buried within the file.
Remove some unnecessary include/using stmts.
Rename pass to MappingInfoCollector.
Rewrite a lot of it so it doesn't use global instance variables and so
it outputs into MappingInfo objects and then dumps those out, instead of going
straight to an assembly file.
Change name of factory to getMappingInfoCollector.
Fold prologue & epilogue writers into MappingInfo methods.
lib/Target/Sparc/FInfo.cpp:
Correct file comment to reflect above change
lib/Target/Sparc/Sparc.cpp:
Change name of factory to getMappingInfoCollector.
llvm-svn: 6617
in Emitter.cpp, just convert the Sparc version of the constant pool into
what's already supported and inter-operate.
* Implemented a first pass at lazy function resolution in the JITResolver. That
required adding a SparcV9CodeEmitter pointer to simplify generating
bit-patterns of the instructions.
* SparcV9CodeEmitter now creates and destroys static TheJITResolver, which makes
sense because the SparcV9CodeEmitter is the only user of TheJITResolver, and
lives for the entire duration of the JIT (via PassManager which lives in VM).
* Changed all return values in the JITResolver to uint64_t because of the 64-bit
Sparc architecture.
* Added a new version of getting the value of a GlobalValue in the
SparcV9CodeEmitter, which now works for already-generated functions (JITted or
library functions).
* Removed little-used and unused functions, cleaning up the internal view of the
SparcV9CodeEmitter.
llvm-svn: 6612
rewriting it. I also vacuumed out all the commented-out code and
inaccurate comments, etc.
(We need to put the mapping information in a data structure so that we can
pass it out to the JIT, instead of automagically converting it to .byte
directives.)
llvm-svn: 6574
they prefer the destination register to be last. Thus, two new classes were made
for them that accomodate for having this layout of operands (F3_1rd, F3_2rd).
llvm-svn: 6564
* Added instruction classes which start building from rs1, then rs2, and rd.
* Fixed order of operands in classes 4.1 and 4.2; added 4.6 .
llvm-svn: 6561
* Added new classes which start building from rs1, adding rs2, and then rd.
* Fixed order of operands in classes 3.11, 3.12, 3.16, and 3.17 .
* Fixed comments to reflect Real Life (tm).
* Removed "don't care" commented out assignments and dead classes (#if 0).
llvm-svn: 6560
SparcInstrSelection.cpp:
* Fixed opcodes to return correct 'i' version since the two functions are each
only used in one place.
* Changed name of function to have an 'i' in the name to signify that they each
return an immediate form of the opcode.
* Added a warning if either of the functions is ever used in a context which
requires a register-version opcode.
SparcV9_F4.td: fixed class F4_3, added F4_4 and notes that F4_{1,2} need fixing
SparcV9.td: added the MOV(F)cc instructions
llvm-svn: 6548
globals in some other places may not have been pulled out either;
globals in phi operands were being put just before the phi instead of
in the predecessor basic blocks.
llvm-svn: 6466
call arguments and return values:
Now all copy operations before and after a call are generated during
selection instead of during register allocation.
The values are copied to virtual registers (or to the stack), but
in the former case these operands are marked with the correct physical
registers according to the calling convention.
Although this complicates scheduling and does not work well with
live range analysis, it simplifies the machine-dependent part of
register allocation.
llvm-svn: 6465
instructions of format 3.12 and 3.13 cannot inherit from F3rdrs1, because that
implies that the two registers are the first two parameters to the instruction.
Thus I made the instructions inherit from F3rd again, and manually added an rs1
field AFTER the shcnt field in the instruction, which maps to the appropriate
place in the instruction.
The other changes are just elimination of unnecessary spaces.
llvm-svn: 6437
SparcV9_F3.td: F3_12 and F3_13 instructions have rd and rs1 fields. Also,
their fields were totally screwed up. This seems to fix the problem.
llvm-svn: 6429
exit code. This, in turn, makes an empty file SparcV9CodeEmitter.inc, and only
much later, produces a link error because the key function that TableGen creates
isn't found.
Using a temporary file in the middle forces a good .INC file to be generated by
TableGen, and it will keep trying until you fix the input file.
llvm-svn: 6392
list (rd, rs1, imm), in that order (bit-wise), the actual assembly syntax is
instr rd, imm, rs1, and that is how they are constructed in the instruction
selector. This fixes the discrepancy.
Also fixed some comments along the same lines and fixed page numbers referring
to where instructions are described in the Sparc manual.
llvm-svn: 6384
Non-obvious change: since I have changed ST and STD to be STF and STDF to
(a) closer resemble their name (NOT assembly text) in the Sparc manual, and
(b) clearly specify that they they are floating-point opcodes,
I made the same changes in this file.
llvm-svn: 6376
Here I had to make one non-trivial change: add a function to get a version of
the opcode that takes an immediate, given an opcode that takes all registers.
This is required because sometimes it is not known at construction time which
opcode is used because opcodes are passed around between functions.
llvm-svn: 6375
annotations on instructions to specify which format they are (i.e., do they take
2 registers and 1 immediate or just 3 registers) as that changes their binary
representation and hence, code emission.
This makes instructions more like how X86 defines them to be. Now, writers of
instruction selection must choose the correct opcode based on what instruction
type they are building, which they already know. Thus, the JIT doesn't have to
do the same work by `discovering' which operands an instruction really has.
As this involves lots of small changes to a lot of files in lib/target/Sparc,
I'll commit them individually because otherwise the diffs will be unreadable.
llvm-svn: 6371
a default value is set in the header file.
* Fixed some code layout to make it more consistent with the rest of codebase
* Added addPassesToJITCompile() with relevant passes
llvm-svn: 6369
Also, added annotations to how instructions are modified (reg/imm operands).
Added prototype for adding register numbers to values pass for interfacing with
the target-independent register allocators in the JIT.
llvm-svn: 6366
that assembles instructions is generated via TableGen (and hence must be built
before building this directory, but that's already the case in the top-level
Makefile).
Also added is .cvsignore to ignore the generated file `SparcV9CodeEmitter.inc',
which is included by SparcV9CodeEmitter.cpp .
llvm-svn: 6357
Fixed spilling of %fcc[0-3] which are part of %fsr.
Moved some machine-independent reg-class code to class TargetRegInfo
from SparcReg{Class,}Info.
llvm-svn: 6339
* Frame information is now stuck in MachineFunctionInfo instead of directly
in MachineFunction.
* Don't require a TM as an argument to the ctor
llvm-svn: 5172
overhead: Merge 3 parallel vectors into 1, change regsUsed hash_set to be a
bitvector. Sped up LLC a little less than 10% in a debug build!
llvm-svn: 4261
interface with code that uses symbols in the ansi-c protected namespace. In
most cases this comes from system header files, such as stdio.h. In particular,
without this change, a reference to the __iob symbol is mangled into ll_iob,
which is not resolved by libc.
llvm-svn: 4165
(1) Padding bytes between structure fields (for alignment) were never
being emitted into the constant pool so the layout did not match!
(2) In printing constants, structures containing structures or arrays
were never handled.
(3) Support new model for external/uninitialized/initialized globals.
Uninitialized globals are no longer emitted since they are external.
Initialized globals may go either in .bss or in .data.
llvm-svn: 4134
(2) Use intelligent multiply selection code for array allocas.
(3) Don't use cache padding for alloca'd stack slots!
(4) Bug fix in handling call arguments: was not copying sixth FP arg
to int reg. when calling a function with no prototype.
llvm-svn: 4130
instructions so that (a) constant folding is done automatically before
code generation, and (b) selection does not have to deal with them.
Also, check for ConstantPointerRefs in additional to GlobalValues
when creating a GEP to load a global address.
llvm-svn: 4126
Don't allow direct FP-to-uint conversion (must be eliminated by preselection).
Address arithmetic for arrays is now entirely 64-bit so no sign-ext needed.
llvm-svn: 3961
Original logs for PreSelection.cpp:
revision 1.2
date: 2002/09/17 23:50:32; author: lattner; state: Exp; lines: +1 -3
Don't put default parameter values into .cpp files, it breaks 3.x compilers
revision 1.1
date: 2002/09/16 15:31:13; author: vadve; state: Exp;
New preselection pass that specializes LLVM code for a target machine,
while remaining in legal portable LLVM form and preserving type
information and type safety.
llvm-svn: 3838
TargetMachine.cpp, since it is entirely machine-independent.
Also, add options to disable optional back-end passes (preselection and
instr. scheduling).
llvm-svn: 3740
-- Add support for ConstantExpr constants (only cast and add operators so far)
-- Avoid generating label Bbss.bss, which sometimes came out twice.
llvm-svn: 3578
SetOperandsForMemInstr significantly). Load and Store
no longer have any indices. Eliminate spurious sign-extension
on a cast to float/double.
llvm-svn: 3498
- Eliminate Sparc*Order classes, moving contents to Sparc*Class classes
- get register name methods now return const char * const pointers instead
of std::string's.
- Added file header comments to SparcRegClassInfo.(cpp|h)
- Moved BIG constant arrays out of SparcRegClassInfo.h, into
SparcRegInfo.cpp. This should allow a LOT of std::string constructors to
not have to be called.
llvm-svn: 3295
and generate actual machine instruction sequences directly.
Also a couple of bug fixes in code for putting constants into registers:
-- Do *not* sign-extend unsigned constant that is shorter than int reg size
-- Fix handling of address constant (a GlobalValue) vs. constant that
must be loaded.
llvm-svn: 2856
-- FP argument to a function with no prototype going on stack
was not being copied to the stack in colorCallArgs().
-- Put caller-saving code *before* argument copying code so that
we don't trash a register before saving it!
-- Two other minor fixes.
llvm-svn: 2855
basic block.
Mark setCCInstr used as dest. of conditional-move as both a def and a use.
BA instruction no longer has the unused CC argument.
llvm-svn: 2836
handle conditional move instructions:
-- cpMem<->Reg functions now support CC registers (int and FP) correctly.
Also, cpMem<->Reg functions now return a vector of machine instructions.
-- Scratch registers must be explicitly provided to cpMem<->Reg when
needed, since CC regs need one to be copied to/from memory.
-- CC regs are saved to a scratch register instead of stack.
-- All regs used by a instruction are now recorded in MachineInstr::regsUsed,
since regs used to save values *across* an instruction are not obvious
either from the operands or from the LiveVar sets.
-- An (explicit or implicit) operand may now be both a def and a use.
This is needed for conditional move operations.
So an operand may need spill code both before and after the instruction.
Other changes:
-- Added several get{Class,Type} functions.
-- Added unified-to-local register number conversion.
-- class MachineCodeForBasicBlock is now an annotation on BasicBlock.
-- Suggest/Color methods may modify the MachineInstr (and always did),
so don't make that argument const!
-- Caller-saving code doesn't need its special purpose code for
handling CC registers since cpMem<->Reg handle those correctly now.
llvm-svn: 2834
-- correct sign extensions for integer casts and for shift-by-constant
instructions generated for integer multiply
-- passing FP arguments to functions with more than 6 arguments
-- passing FP arguments to varargs functions
-- passing FP arguments to functions with no prototypes
-- incorrect stack frame size when padding a section
-- folding getelementptr operations with mixed array and struct indexes
-- use uint64_t instead of uint for constant offsets in mem operands
-- incorrect coloring for CC registers (both int and FP): interferences
were being completely ignored for int CC and were considered but no
spills were marked for fp CC!
Also some code improvements:
-- better interface to generating machine instr for common cases
(many places still need to be updated to use this interface)
-- annotations on MachineInstr to communicate information from
one codegen phase to another (now used to pass information about
CALL/JMPLCALL operands from selection to register allocation)
-- all sizes and offests in class TargetData are uint64_t instead of uint
llvm-svn: 2640
- Rename runOnMethod to runOnFunction
* Transform getAnalysisUsageInfo into getAnalysisUsage
- Method is now const
- It now takes one AnalysisUsage object to fill in instead of 3 vectors
to fill in
- Pass's now specify which other passes they _preserve_ not which ones
they modify (be conservative!)
- A pass can specify that it preserves all analyses (because it never
modifies the underlying program)
* s/Method/Function/g in other random places as well
llvm-svn: 2333
Vikram S. Adve